deep learning

In this paper, we propose an automatic detection system for COVID-19 cases based on the Internet of Things. In the proposed model, first, using Internet of Things technology, medical images are sent directly to the data collection after the suspicious person's visit through medical equipment equipped with Internet of Things, and then, in order to help radiologists to interpret medical images better, usage has been made of four pre-trained convolutional neural network models i.e. InceptionV3, InceptionResNetV2, VGG19 and ResNet152 as well as two datasets of chest radiology medical images and CT Scan in a 3-class classification for accurate prediction of cases suffering from COVID-19, healthy people, and diseased cases. Finally, the best result for CT-Scan images is related to InceptionResNetV2 architecture with an accuracy of 99.366%, and for radiology images related to the InceptionV3 architecture, it is 96.943%. The results show that this system leads to a reduction in daily visits to medical centers and thus reduces the pressure on the medical care system. It also helps rheology specialists to identify the disease as quickly as possible.

This paper introduces an efficient intrusion detection system for the Internet of Things, addressing the challenge of malware-infected IoT nodes acting as botnet attackers, along with issues in existing intrusion detection systems such as feature selection, data imbalance, and centralization. The proposed system leverages the distributed architecture of SDN. The method begins by balancing the dataset using the SMOTE technique. Essential features are then selected using the African Vulture Optimization Algorithm. Subsequently, an LSTM deep learning model is trained within the SDN controller. SDN switches utilize this trained model for attack detection. To enhance attack mitigation, attacking node addresses are shared among SDN switches, ensuring consistent recognition and enabling effective Distributed Denial-of-Service (DDoS) attack prevention across the network. Experimental results obtained in MATLAB, using the NSL-KDD dataset, demonstrate the proposed method’s effectiveness, achieving an accuracy of 99.34%, a sensitivity of 99.16%, and a precision of 98.93% in attack detection. The proposed method outperforms feature selection methods based on WOA, HHO, and AO algorithms, and deep learning methods like LSTM, RNN, and CNN, particularly in detecting DDoS attacks.

Enhancing medical devices with Internet of Things (IOT) and diagnostic artificial intelligence technology, while considering the constraints of these systems, has the potential to modernize and enhance the diagnostic approach of future generations of Internet of Things systems in healthcare. The radiography device, commonly used in paraclinical settings, is widely used in various hospital departments. The automated assessment of abnormalities and bone age from radiographic images of the left hand assists radiologists, pediatricians, and forensic experts in determining the developmental stage of young individuals. The IoT devices in medicine are unable to process large amounts of data due to limited resources. This article uses a teacher-student network for bone age classification, using the decomposed knowledge distillation model of convolutional neural networks. This approach minimizes the computational resources needed for edge devices. The proposed method is comprised of two sequential steps.  In the preprocessing step, the initial phase involves the elimination of non-clinical data and artifacts. This is followed by the extraction of region of interest (ROI). In this phase of the procedure, only the hand portion of the patient's X-ray remains for further evaluation. The subsequent phase involves the delineation of the boundaries of the region of interest. This is necessary because, in certain age groups, some bones are not ossified. Consequently, reliance on bones as landmarks is precluded.  In the second step, The extracted ROI from the preceding step is utilized to train the teacher model. The student model utilizes the teacher model's knowledge to learn how to predict patient age. Therefore, the present study puts forth transfer learning methodologies founded on the distillation of knowledge, with the aim of facilitating the transference of knowledge between teacher and student models. The proposed method is based on the data set of the Digital Hand Atlas (DHA) database. The evaluation criteria used in this work are Accuracy, recall, permission and mean absolute error (MAE). The proposed model achieves 96/47% test accuracy for bone age classification.

Deep learning is an essential tool for detecting distributed denial of service (DDoS) attacks due to its ability to analyze complex network traffic patterns and respond in real-time. However, a comprehensive examination of the opportunities and challenges in this field is necessary, given its emerging nature. This examination should include real implementations or benchmark data samples. In this paper, we introduce two methods for detecting DDoS attacks: one using Long Short-Term Memory (LSTM) and the other using Convolutional Neural Networks (CNN). Additionally, we propose a new method that combines LSTM and CNN. The results demonstrate that both LSTM and LSTM-CNN methods consistently outperform CNN in terms of accuracy, precision, recovery, and F1 scores. Our investigations reveal that CNN can automatically learn features such as packet size, timing, and source/destination addresses from raw network traffic. On the other hand, LSTM is particularly useful for detecting temporal sequence patterns of attacks in network traffic. The choice between CNN or LSTM for DDoS detection depends on the specific characteristics of the attack dataset and the relative importance of spatial and temporal features in identifying DDoS attacks. Finally, we examine challenges such as overfitting, computational complexity, interpretability, data limitations, and hostile attacks. Doubts surrounding the reporting of results in literature can be attributed to problems with the benchmark dataset used, including limited sample size and variety, lack of labeling, and unbalanced data.

Graph structures play a vital role in modeling relationships across various domains, including social networks, knowledge bases, and biological networks. As the dimensions of these networks grow, the efficiency of proximity-based analysis methods declines, necessitating the use of graph embedding techniques to reduce dimensionality while preserving the underlying structure. This process enhances performance in applications such as node classification and link prediction. However, traditional graph embedding methods face challenges with capturing non-linear relationships and scaling to large networks. Additionally, in real-world networks, the essential initial and precise node features which are required by these algorithms are not always available. In this paper, we propose a novel framework called FuzzyRandomNet, which addresses these challenges by integrating fuzzy logic with random walks. FuzzyRandomNet introduces non-linear layers and optimizes node features to provide more efficient and scalable solutions for graph representation learning. The evaluation of the proposed method against existing techniques on standard datasets demonstrates superior performance in node classification and link prediction, exhibiting higher accuracy and flexibility in large and complex networks.

In today's society, news and advertisements have a special place in the growth and development of society. By specifying the main words of the ad, you can understand its general meaning. Preparing these words in the traditional way requires time and specialized knowledge about the subject of the text. Ideakav site is a system that collects Telegram messages and advertisements. The requirement of the idea search system was to extract keywords from the advertisements published in Telegram. The quality of extracted keywords plays a significant role in improving SEO and advertising statistics. By using embedding algorithms, it is possible to extract colloquial conversations and the semantic structure of the text, therefore, it is useful in identifying keywords in Telegram ads that are often published in popular form. In this research, a model of word embedding has been implemented using the data of the idea mining system. The innovation used in this research is created by combining word embedding methods, word frequency and word position. The embedding model is created from two-word words. Creating a model of two-word words is because most of the keywords consist of two words or more. In order to better display the evaluations, the IK model (proposed model) has been compared with statistical methods and graph-based methods, and the obtained results show that the combination of the two-gram IK model has produced a better performance in extracting keywords than other methods.

This paper presents a compact neural architecture search method for image classification using the Gravitational Search Algorithm (GSA). Deep learning, through multi-layer computational models, enables automatic feature extraction from raw data at various levels of abstraction, playing a key role in complex tasks such as image classification. Neural Architecture Search (NAS), which automatically discovers new architectures for Convolutional Neural Networks (CNNs), faces challenges such as high computational complexity and costs. To address these issues, a GSA-based approach has been developed, employing a bi-level variable-length optimization technique to design both micro and macro architectures of CNNs. This approach, leveraging a compact search space and modified convolutional bottlenecks, demonstrates superior performance compared to state-of-the-art methods. Experimental results on CIFAR-10, CIFAR-100, and ImageNet datasets reveal that the proposed method achieves a classification accuracy of 98.48% with a search cost of 1.05 GPU days, outperforming existing algorithms in terms of accuracy, search efficiency, and architectural complexity.

Improving image quality for identification and authentication in security and surveillance systems is of particular importance, and today, using artificial intelligence, the quality of images can be significantly improved. In this regard, the present paper, focusing on the details of face images, has improved the image failure detection model in the adversarial generator network, which led to a suitable performance in the meta-dissolving of face images. Most of the CNN networks that have been presented in recent years require a large set of images with appropriate annotations for proper performance, and they usually perform poorly in the case of degradation that have not been trained, which is addressed in this research to improve this challenge. In this work, pairs of high-quality and low-quality images are used to train the image degradation detection model; This information is then transferred to real images. The naturalness of the output images is one of the most important challenges in this field. The obtained results show that the criterion of perceptual similarity of the obtained image is equal to 38.4%, which is comparable to recent researches. As a result, using the proposed model, more natural images were produced